Preserving Disagreement: Architectural Heterogeneity and Coherence Validation in Multi-Agent Policy Simulation

1. Problem Statement

The paper addresses a critical failure mode in Large Language Model (LLM) based multi-agent deliberation systems used for normative policy simulation: Artificial Consensus.

• The Phenomenon: In policy deliberation (where no objectively correct answer exists), agents assigned different value perspectives (e.g., Security vs. Risk Tolerance) often converge on the same policy option. This creates an illusion of agreement that obscures genuine trade-offs and value tensions.
• Causes:
  1. Shared Inductive Bias: When all agents use the same underlying model, they share training data, RLHF preferences, and reasoning patterns, effectively acting as the same reasoner wearing different "hats."
  2. Debate Capture: Minority perspective agents are persuaded by majority arguments during debate phases, abandoning their assigned viewpoints to adopt the most compelling (often majority) position.
• Distinction from Accuracy Tasks: Unlike factual or mathematical debates where convergence signals correctness, in normative deliberation, convergence is a failure mode. The system's value lies in mapping divergence, not resolving it.

2. Methodology: The AI Council Framework

The authors propose The AI Council, a three-phase deliberation system designed to simulate multi-stakeholder policy debates while preserving value-based disagreement.

System Architecture

The system runs on consumer hardware using seven locally hosted 7–9B parameter models, with a single API call to a frontier model for validation.

1. Phase 1: Structured Debate: Three "Champion" agents advocate for specific policy options. They present position papers, critique other options from their specific value lens, and defend their positions.
2. Phase 2: Independent Evaluation: Seven "Evaluator" agents, each primed with a distinct value perspective (e.g., Security Focus, Pragmatism, Risk Tolerance), independently read the debate corpus and rank the options. Crucially, evaluators do not see each other's outputs.
3. Phase 3: Coherence Validation (Optional): A "Delphi" layer uses a frontier model (Claude Sonnet 4) to assess whether each evaluator's reasoning is genuinely grounded in their assigned value perspective, generating a coherence score [0.0, 1.0].

Experimental Design

The study employs a three-state design across two policy scenarios:

• State A (Homogeneous): All agents use the same base model (Mistral-7B).
• State B (Heterogeneous): Each agent uses a different model from a diverse pool (7–9B models from various organizations, e.g., Qwen, Mistral, DeepSeek, Gemma).
• State C (Heterogeneous + Delphi): State B runs post-processed with coherence-weighted scoring (using a modified Borda count).

Scenarios Tested:

1. Child Welfare Intervention: Features a "dominant option" (Intensive family preservation) where one choice is inherently superior, testing the system's ability to detect dominance vs. artificial consensus.
2. Urban Housing Crisis: Features "genuinely competitive options" (Rent Control vs. Deregulation vs. Public Housing) with no inherent hierarchy, testing the system's ability to maintain diverse viewpoints.

Metrics:

• First-Choice Concentration (FCC): Measures the degree of agreement on the top option.
• Borda Margin: The gap between the top two ranked options.
• Effective Perspectives: Shannon entropy of the vote distribution.
• Trustworthy Tension Rate: The fraction of theoretical value conflicts where both sides reason faithfully (high coherence).

3. Key Contributions

1. Architectural Heterogeneity as a Solution: Demonstrates that assigning different LLMs to different value perspectives significantly reduces artificial consensus in normative tasks, a finding that contrasts with accuracy-oriented debate literature.
2. The Fidelity–Diversity Tradeoff: Introduces and documents a novel tradeoff where assessing reasoning quality (fidelity) can inadvertently reduce diversity if high-quality models happen to agree, while lower-quality models provide the dissenting voices.
3. Empirical Constraints on Small Models: Reports negative results from three failed Delphi designs (peer exposure, value-tension filtering, self-confrontation), revealing that 8B models exhibit binary responses to counter-arguments (either maintaining position or capitulating entirely) rather than graded nuance.
4. Diagnostic Metric: Proposes the Trustworthy Tension Rate as a baseline measure for the capability of small-model deliberation systems.

4. Key Results

A. Architectural Heterogeneity (State A vs. State B)

Using different models for different perspectives significantly reduced artificial consensus in both scenarios:

• Child Welfare: First-choice concentration dropped from 70.9% (Homogeneous) to 46.1% (Heterogeneous) ($p < 0.001$, $r=0.58$).
• Housing: First-choice concentration dropped from 46.0% to 22.9% ($p < 0.001$, $r=0.50$).
• Interpretation: Heterogeneity disrupts shared inductive biases. Unlike accuracy tasks where models converge on the "truth" regardless of architecture, in normative tasks, different architectures encode different value priors, preventing spurious agreement.

B. Coherence Validation (State B vs. State C)

The effect of coherence validation was scenario-dependent, revealing the Fidelity–Diversity Tradeoff:

• Child Welfare (Dominant Option): Coherence validation reduced concentration further (46.1% $\to$ 40.8%). It successfully downweighted low-coherence evaluators who had been "captured" by the majority view, amplifying the few genuine dissenters.
• Housing (Competitive Options): Coherence validation increased concentration (22.9% $\to$ 26.6%). In this scenario, the highest-coherence models (assigned to Performance Focus and Risk Tolerance) naturally clustered on the "Deregulate" option. The low-coherence models (Pragmatism, Simplicity) were the ones providing diversity. Weighting by fidelity inadvertently silenced the diverse, lower-quality voices, increasing consensus.

C. Model Behavior & Negative Results

• Binary Response: 8B models failed to engage in "consider and reject" reasoning. When exposed to counter-arguments in failed Delphi rounds, they either held firm or completely flipped (sycophantic capitulation).
• Trustworthy Tension Rate: Across both scenarios, only ~50% of theoretical value conflicts were "trustworthy" (both sides reasoning faithfully). This establishes a baseline limitation for current 7–9B model deliberation capabilities.

5. Significance and Implications

• For Policy Simulation: The paper argues that model diversity is not optional but a primary mechanism for generating useful outputs in normative domains. Relying on a single model (even with different prompts) leads to artificial consensus.
• For AI Safety: The findings highlight that small models (7–9B) are fragile under adversarial debate conditions, exhibiting binary capitulation. Architectural isolation (preventing agents from seeing each other's outputs during evaluation) is a necessary safeguard.
• Design Principle: The Fidelity–Diversity Tradeoff suggests that in multi-agent systems, quality weighting (fidelity) and diversity preservation are competing objectives. Practitioners should report both weighted and unweighted results; the divergence between them indicates which perspectives are being modeled reliably and which are suffering from low-fidelity reasoning.
• Limitations: The study relies on a specific pool of 7–9B models and two scenarios. The "Trustworthy Tension Rate" of ~50% indicates that while the system preserves disagreement, the quality of that disagreement is currently limited.

In conclusion, the AI Council demonstrates that preserving genuine disagreement in AI policy simulation requires architectural heterogeneity to break shared biases, while acknowledging that coherence validation introduces complex trade-offs between reasoning quality and viewpoint diversity.